Understanding the Dynamics of Pricing and Negotiation

1
Simulating a Network of Trading Agents

• Understanding the Dynamics of Pricing and
  Negotiation

Presented by Rodolfo Garcia-Flores and Nectarios
Kontoleon 5th Annual CSIRO CSS Workshop Sydney,
8th to 10th of August, 2006
2
Contents

• Motivation
• The Colour Trading World
• Implementation
• Alternative Strategies
• Optimal
• Heuristic (Simulated Annealing)
• Price Functions
• Measuring Performance
• Preliminary Results
• Directions for future work

3
Motivation

• Commercial negotiations in the near future will
  be performed by proxy agents over the Internet.
• Negotiations will involve many parties in a
  network of economically motivated entities.
• Agents will have access to limited information.
• Performance will depend on strategy.
• Topology of the network is another important
  factor.
• It is imperative to understand the economic
  incentives and behaviours in order to anticipate
  their collective interactions.

4
Motivation

• A model world is used to explore trading agents
  in a network.
• Knowledge related to emergent behaviour,
  connectivity, strategies and network dynamics
  will feed the groups applied research and
  flagship projects.

5
Relevance for Strategic Research

• Partnership research to move from linear value
  (supply) chains to adaptive networks.
• Partners of the network collaboratively adapt to
  the changes in their shared environment.
• CMIS attempts to develop decision technologies
  where there is appreciable diversity in
  stakeholder capabilities and goals.
• High-impact RD targets are those in which the
  supply network must address a particularly
  complex activity coordination requirement.

6
We want to explore

• How can we predict system performance from agent
  strategies and network topology?
• How does the best strategy depend on position in
  network?
• How does the profit of an agent with a certain
  strategy depend on the strategies of other
  agents?
• How does the individual optimal strategies affect
  the overall system optimum?
• Emergent behaviour when there is a trade-off
  between control and flexibility.

7
The Colour Trading World

• Sources
• Produce units of colour
• Have fixed production costs per unit
• Sinks
• Consume colour
• Get more profit for selling bundles of different
  colour in fixed ratio
• Merchants
• Trade colour
• Make profit by buying cheap and selling dear

8
Complexity of the Sinks

• Sinks consume colour in bundles
• Each bundle is defined by
• A vector v containing an integer vc for each
  colour c
• A revenue r per unit consumed
• A maximum number m of units that can be consumed
• One unit of the bundle
• contains vc units of colour c
• Brings revenue of r
• The sink can consume at most m units of the bundle

9
Bundle Example for a Sink

• Existing colours red, green, blue
• Bundles of sink t
• B1 V (1,0,0), r 2, m 3
• B2 V (0,1,0), r 1, m 4
• B3 V (0,0,1), r 3, m 2
• B4 V (1,2,1), r 10, m 5
• B5 V (2,1,1), r 15, m 3
• Optimal colour distributions on bundles using
  linear program
• Agents use heuristics in implementation

10
The Colour Trading World
11
The Colour Trading World
(55)
(23)
(44)
(42)
(56)
(22)
(35)
12
The Colour Trading World

• Rules of the game
• Agents only know their direct neighbours.
• Agent negotiate contracts for delivery of amounts
  of flow per time unit.
• Agents are not allowed to enter into a contract
  that might result in a loss.
• If any offer or request an agent has sent out is
  accepted, the agent must be able to meet the
  corresponding obligation.

13
The Colour Trading World

• Offers and requests for colour are valid until
  they are cancelled.
• Game proceeds in rounds.
• In each round, each agent gets a turn for sending
  out offers, requests, accepts and cancels.
• Every accept message is honoured.
• Reactions to accepts and cancels are propagated
  through the network before the next agent's turn
  starts.
• Game ends when agents run out of new things to
  say.

14
The Colour Trading World
15
The Colour Trading World
Has blue (53pu)
16
The Colour Trading World
Has blue (53pu)
17
The Colour Trading World
Has blue (53pu)
55pu
55pu
18
The Colour Trading World
Has blue (53pu)
55pu
56pu
55pu
19
The Colour Trading World
Has blue (53pu)
55pu
56pu
57pu
55pu
20
The Colour Trading World
Sink accepts
Has blue (53pu)
55pu
56pu
57pu
55pu
21
The Colour Trading World
Sink accepts
Has no blue
54pu
55pu
56pu
57pu
55pu
22
Our Colour Trading Implementation

• Agents based system using JADE framework
• Each agent knows its neighbours and its successor
  in the turn
• When it's an agent's turn, it sends out offers,
  requests, accepts or cancels to its neighbours
• When agent has finished sending its messages, it
  sends the your turn message to its successor
• Reception of accepts or cancels by agent triggers
  consistency reactions
• Agent reacts immeatedly, even when it's not its
  turn
• Consistency reactions involve updating internal
  data and maybe also sending out accepts and
  cancels

23
Structure of the Implementation
Java
Jade
Colour Agent

• Inter-Agent Communication
• Thread Control

FSM-Behaviour
Communication Protocol
Agent Brain
Strategy
24
Structure of the Implementation
Java
Jade
Colour Agent

• Inter-Agent Communication
• Thread Control

FSM-Behaviour
Communication Protocol
Source Brain
Source Strategy
25
Structure of the Implementation
Java
Jade
Colour Agent

• Inter-Agent Communication
• Thread Control

FSM-Behaviour
Communication Protocol
Sink Brain
Sink Strategy
26
Structure of the Implementation
Java
Jade
Colour Agent

• Inter-Agent Communication
• Thread Control

FSM-Behaviour
Communication Protocol
Merchant Brain
Merchant Strategy
27
Structure of the Implementation
Java
Jade
Colour Agent

• Inter-Agent Communication
• Thread Control

FSM-Behaviour
Communication Protocol
Merchant Brain
Simple Merchant Strategy
28
Structure of the Implementation
Java
Jade
Colour Agent

• Inter-Agent Communication
• Thread Control

FSM-Behaviour
Communication Protocol
Merchant Brain
Sophisticated Merchant Strategy
29
Strategies

• All agents types can potentially use complex
  strategies in negotiations with neighbours
• Additional merchant complexity
• Merchants could aggregate and split supply and
  demand to formulate offers and requests
• Deciding which messages to accept becomes a hard
  problem (subset sum problem)
• Additional sink complexity
• Sinks can make profits by obtaining bundles of
  different colours in specific ratios
• Problem of deciding how best do distribute
  obtained colours into bundles can be solved using
  a linear program

30
Currently Used Strategies

• Simplest strategies that could possibly work
• All agents calculate prices in their messages to
  guarantee minimum profit of one currency unit
  profit for every unit of colour moved
• First offers are final offers
• Sources tell all their neighbours what they
  offer, update offers when sale happens
• Merchants tell all neighbours about offers and
  requests they know
• Sinks ask all their neighbours for colour they
  need, update requests when sale happens
• All agents accept incoming messages if they meet
  minimum profit criterion

31
A Strategy Test-bed
32
A Strategy Test-bed

• Agents strive to maximise their profit subject to
  a production function.
• Markets simply determine equilibrium prices and
  quantities for agents trading a given good.
• Agents in the extreme tiers of the market are
  connected to external supply and demand functions
  of the form
• Where I/O are input/output, p/q are
  buying/selling prices, and a,b,c,d are constant
  parameters.

33
A Model of Behaviour

• Maximise profit, i.e.,
• Subject to

34
Optimal and Heuristic Strategies
35
A Simple GUI
36
Preliminary Results
37
Preliminary Results
38
Preliminary Results
39
Preliminary Results
40
Aggregate Price Functions

• Merchant has received two offers
• Up to 5 units of blue for 3 per unit
• Up to 7 units of blue for 5 per unit

41
Aggregate Price Functions

• Merchant has received two offers
• Up to 5 units of blue for 3 per unit
• Up to 7 units of blue for 5 per unit
• Resulting unit price curve

42
Consequences of Aggregation

• Price per unit functions can become very complex
• Messages should be able to describe price per
  unit functions
• Agents want to find price/amount where they make
  maximum profit
• Merchants can act in similar way as an electronic
  stock market

43
Measuring Performance

• Optimal profit of whole network can be computed
  with linear program (LP) if fractional units of
  colour are allowed
• Agents have only local knowledge while LP knows
  complete network gt expect lower profit for agent
  system
• Other Preformance measure is number of rounds
  before all trades are executed

44
Project Future

• Try more advanced strategies
• Examine interplay of different strategies
• Does a clever agent in a network of stupid agents
  only improve its own profit or also the
  performance of the whole network?
• Should agents be accomodating or single mindedly
  egoistic to achieve maximum network performance?
• How much can a clever agent profit from knowing
  more about its position in the network?

45
Contact Information
For more information, contact